DeepSeek has put a working sketch of its super-agent idea on GitHub, and the first thing the README does is lower expectations.
What it is. According to the dsh-superagent repository, the project is a proactive coding agent that serves a single GitHub repository. It ships as a bundle for DeepSeek Harness. It listens to GitHub and can optionally listen to Feishu as well.
How a job flows. When a person the team has allowed asks for something, the agent opens a work item. It handles that item in its own agent session and its own directory, then replies in the conversation where the request started. A reply can be a pull request. A web panel lists what needs a person’s attention.
The design rule. The README frames the project as a way of working, not only as code. The goal is to spend as little human attention as possible while delivering as much value as possible. Because an agent acting for a team can also waste that team’s attention, it acts only when someone asks or when a rule the team wrote permits it. It speaks up only when a person has to act or can use a result.
Setup by agent. Installation has many steps, so the README tells you to hand it to your own coding agent and point it at SETUP.md. Your agent then asks brief questions one by one, presents the whole plan a single time, and handles installation and configuration.
The fine print. The repo calls itself a research prototype for reference only, not an official DeepSeek product. The team says it grew out of their own daily work, with coding agents doing much of the engineering. It should be treated as research code: it may have bugs, its style may be uneven, and you should review and adapt it before use.
Not the full vision. The published code was adapted for release. It is not identical to the internal version, or to the system described in the article Exploring Super-agent Design: Towards a More Autonomous Agent System. Some ideas in that article are future directions that this prototype does not implement.
The basics. The code is under the MIT license. The README also points readers to a set of numbered documents, from philosophy and domain model through authorization, execution, prompts, the panel, deployment, architecture and a cookbook, meant to be read in order. For development, running pnpm check handles the typecheck, the tests and the build.
What’s real here: a candid research prototype, released for reference only, of an agent that acts only when someone asks or a team rule permits it. Read it for the ideas, and follow the README’s own advice before running it on anything that matters.
Sources
By Willa Ravenwire · The AI Bandwagon
